RNA sequencing-based deep learning techniques to predict response to immune checkpoint inhibitors in metastatic renal cell cancer patients

Immune checkpoint inhibitors have become a cornerstone in the management of metastatic Renal Cell Carcinoma (mRCC). However, to identify the most suitable patients for this treatment is an unmet need. We aimed to explore the utility of Deep Learning (DL) techniques using transcriptomic data to predict response to immunotherapy. We conducted a retrospective analysis using publicly available data from patients treated with nivolumab in the clinical trials Chekmate 009, 010 and 025. The primary objective was to assess the performance of different DL models such as Deep Autoencoder (DA) and Convolutional Neural Networks predicting the Progression Free Survival (PFS) of these patients. With that scope, we followed several research lines including the creation of combined datasets with clinical and RNA sequencing data and comparing the results of the DL models against the performance of traditional Machine Learning (ML) models. Finally, we came up with an interpretability analysis of the most accurate models. Clinical and transcriptomic data from 181 nivolumab-treated patients were analyzed, confirming the feasibility of modeling immunotherapy response with RNA sequencing data. However, DL models did not significantly outperform traditional ML approaches in response prediction. Although DA provided 65 % F1-score, the multi-layer perceptron classifier emerged as the most effective, achieving an 75 % F1-score in binary PFS classification, employing a selection RNA features and clinical ones. Moreover, the integration of clinical and molecular data yielded more accurate results than any dataset by its own. In conclusion, using traditional ML algorithms and the combination of clinical and molecular data led to the most accurate predictions of mRCC treatment response. These findings underscore the necessity for continued exploration in DL analysis, considering potential biases due to data codification. • Based on 15 transcripts and 8 clinical variables, the multilayer perceptron reached a hit rate of 75 %. • The model identified patients likely to respond to nivolumab versus those needing to be offered alternative treatments. • DL models were not significantly better than traditional ML due to high dimensionality and noise. • Interpretability results revealed that RNA features are the most valuable variables for the decision making. • Combining clinical and molecular data with traditional ML gave the most accurate mRCC predictions.